Hongtao Mu

Recycling just four transformer layers through twelve recurrent iterations matches full-depth baselines on core linguistic benchmarks, exposing precisely where compute can—and cannot—substitute for raw parameter capacity.